Adjusting for Selection Bias Due to Missing Data in Electronic Health Records-Based Research by Blending Multiple...

Area of Science:

• Health Informatics
• Biostatistics
• Observational Studies

Background:

• Missing data is a significant challenge in large observational studies using electronic health records (EHR).
• Standard methods for handling missing data in EHR often fail to account for the complex, heterogeneous data structure.
• Existing frameworks may not adequately address selection bias introduced by incomplete data.

Purpose of the Study:

• To formalize a previously proposed framework for analyzing EHR data with missingness.
• To propose a pragmatic, flexible, and scalable framework for estimation and inference in EHR studies.
• To improve the alignment between missingness assumptions and the complexity inherent in EHR data.

Main Methods:

• Development of a novel framework blending inverse probability weighting and multiple imputation.
• Formalization of analyses within a framework that modularizes data provenance as a sequence of decisions.
• Application of the framework to investigate weight loss outcomes following bariatric surgery using EHR data.

Main Results:

• The proposed framework offers a more robust approach to handling missing data in EHR.
• Demonstrated improved alignment between missingness assumptions and EHR data complexity.
• Illustrated the framework's utility in a real-world data application concerning bariatric surgery outcomes.

Conclusions:

• The new framework provides a flexible and scalable solution for estimation and inference in EHR studies with missing data.
• This approach enhances the reliability of findings from observational studies utilizing complex EHR data.
• The methodology is applicable to various research questions, including the investigation of treatment effects and effect modification.